I'd like to know if they fed their machine with real arcade tracks or used a different method to train it, like their Pathways Autoregressive Text-to-Image model.
(Parti model avoids visual copyright issues, although the music industry is far more aggressive than visual arts!)